# Copyright 2021 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PyTorch M2M100 model.""" import math from collections.abc import Callable import torch from torch import nn from torch.nn import CrossEntropyLoss from ... import initialization as init from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache from ...generation import GenerationMixin from ...integrations.deepspeed import is_deepspeed_zero3_enabled from ...integrations.fsdp import is_fsdp_managed_module from ...masking_utils import create_bidirectional_mask, create_causal_mask from ...modeling_flash_attention_utils import ( FlashAttentionKwargs, ) from ...modeling_layers import GradientCheckpointingLayer from ...modeling_outputs import ( BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions, Seq2SeqLMOutput, Seq2SeqModelOutput, ) from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import TransformersKwargs, auto_docstring, is_torchdynamo_compiling, logging from .configuration_m2m_100 import M2M100Config logger = logging.get_logger(__name__) # Copied from transformers.models.bart.modeling_bart.shift_tokens_right def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int): """ Shift input ids one token to the right. """ shifted_input_ids = input_ids.new_zeros(input_ids.shape) shifted_input_ids[:, 1:] = input_ids[:, :-1].clone() shifted_input_ids[:, 0] = decoder_start_token_id if pad_token_id is None: raise ValueError("self.model.config.pad_token_id has to be defined.") # replace possible -100 values in labels by `pad_token_id` shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id) return shifted_input_ids # Copied from transformers.models.bart.modeling_bart.BartScaledWordEmbedding with Bart->M2M100 class M2M100ScaledWordEmbedding(nn.Embedding): """ This module overrides nn.Embeddings' forward by multiplying with embeddings scale. """ def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int, embed_scale: float | None = 1.0): super().__init__(num_embeddings, embedding_dim, padding_idx) self.embed_scale = embed_scale def forward(self, input_ids: torch.Tensor): return super().forward(input_ids) * self.embed_scale class M2M100SinusoidalPositionalEmbedding(nn.Module): """This module produces sinusoidal positional embeddings of any length.""" def __init__(self, num_positions: int, embedding_dim: int, padding_idx: int | None = None): super().__init__() self.offset = 2 self.num_positions = num_positions self.embedding_dim = embedding_dim self.padding_idx = padding_idx self.make_weights(num_positions + self.offset, embedding_dim, padding_idx) def make_weights(self, num_embeddings: int, embedding_dim: int, padding_idx: int | None = None): emb_weights = self.get_embedding(num_embeddings, embedding_dim, padding_idx) if hasattr(self, "weights"): # in forward put the weights on the correct dtype and device of the param emb_weights = emb_weights.to(dtype=self.weights.dtype, device=self.weights.device) self.register_buffer("weights", emb_weights, persistent=False) @staticmethod def get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: int | None = None): """ Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of "Attention Is All You Need". """ half_dim = embedding_dim // 2 emb = math.log(10000) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, dtype=torch.int64).float() * -emb) emb = torch.arange(num_embeddings, dtype=torch.int64).float().unsqueeze(1) * emb.unsqueeze(0) emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1) if embedding_dim % 2 == 1: # zero pad emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1) if padding_idx is not None: emb[padding_idx, :] = 0 return emb.to(torch.get_default_dtype()) @torch.no_grad() def forward( self, input_ids: torch.Tensor | None = None, inputs_embeds: torch.Tensor | None = None, past_key_values_length: int = 0, ): if input_ids is not None: bsz, seq_len = input_ids.size() # Create the position ids from the input token ids. Any padded tokens remain padded. position_ids = self.create_position_ids_from_input_ids( input_ids, self.padding_idx, past_key_values_length ).to(input_ids.device) else: bsz, seq_len = inputs_embeds.size()[:-1] position_ids = self.create_position_ids_from_inputs_embeds( inputs_embeds, past_key_values_length, self.padding_idx ) # expand embeddings if needed max_pos = self.padding_idx + 1 + seq_len + past_key_values_length if max_pos > self.weights.size(0): self.make_weights(max_pos + self.offset, self.embedding_dim, self.padding_idx) return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, self.weights.shape[-1]).detach() @staticmethod def create_position_ids_from_inputs_embeds(inputs_embeds, past_key_values_length, padding_idx): """ We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids. Args: inputs_embeds: torch.Tensor Returns: torch.Tensor """ input_shape = inputs_embeds.size()[:-1] sequence_length = input_shape[1] position_ids = torch.arange( padding_idx + 1, sequence_length + padding_idx + 1, dtype=torch.long, device=inputs_embeds.device ) return position_ids.unsqueeze(0).expand(input_shape).contiguous() + past_key_values_length @staticmethod # Copied from transformers.models.roberta.modeling_roberta.RobertaEmbeddings.create_position_ids_from_input_ids def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0): """ Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. This is modified from fairseq's `utils.make_positions`. Args: x: torch.Tensor x: Returns: torch.Tensor """ # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA. mask = input_ids.ne(padding_idx).int() incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask return incremental_indices.long() + padding_idx # Copied from transformers.models.bert.modeling_bert.eager_attention_forward def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: torch.Tensor | None, scaling: float | None = None, dropout: float = 0.0, **kwargs: Unpack[TransformersKwargs], ): if scaling is None: scaling = query.size(-1) ** -0.5 # Take the dot product between "query" and "key" to get the raw attention scores. attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling if attention_mask is not None: attn_weights = attn_weights + attention_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1) attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value) attn_output = attn_output.transpose(1, 2).contiguous() return attn_output, attn_weights # Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->M2M100 class M2M100Attention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__( self, embed_dim: int, num_heads: int, dropout: float = 0.0, is_decoder: bool = False, bias: bool = True, is_causal: bool = False, config: M2M100Config | None = None, layer_idx: int | None = None, ): super().__init__() self.embed_dim = embed_dim self.num_heads = num_heads self.dropout = dropout self.head_dim = embed_dim // num_heads self.config = config if (self.head_dim * num_heads) != self.embed_dim: raise ValueError( f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}" f" and `num_heads`: {num_heads})." ) self.scaling = self.head_dim**-0.5 self.is_decoder = is_decoder self.is_causal = is_causal self.layer_idx = layer_idx if layer_idx is None and self.is_decoder: logger.warning_once( f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and " "will lead to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` " "when creating this class." ) self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias) self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias) self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias) self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) def forward( self, hidden_states: torch.Tensor, key_value_states: torch.Tensor | None = None, past_key_values: Cache | None = None, attention_mask: torch.Tensor | None = None, output_attentions: bool = False, cache_position: torch.Tensor | None = None, # TODO: we need a refactor so that the different attention modules can get their specific kwargs # ATM, we have mixed things encoder, decoder, and encoder-decoder attn **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]: """Input shape: Batch x Time x Channel""" # if key_value_states are provided this layer is used as a cross-attention layer # for the decoder is_cross_attention = key_value_states is not None # determine input shapes bsz, tgt_len = hidden_states.shape[:-1] src_len = key_value_states.shape[1] if is_cross_attention else tgt_len q_input_shape = (bsz, tgt_len, -1, self.head_dim) kv_input_shape = (bsz, src_len, -1, self.head_dim) # get query proj query_states = self.q_proj(hidden_states).view(*q_input_shape).transpose(1, 2) is_updated = False if past_key_values is not None: if isinstance(past_key_values, EncoderDecoderCache): is_updated = past_key_values.is_updated.get(self.layer_idx) if is_cross_attention: # after the first generated id, we can subsequently re-use all key/value_states from cache curr_past_key_values = past_key_values.cross_attention_cache else: curr_past_key_values = past_key_values.self_attention_cache else: curr_past_key_values = past_key_values current_states = key_value_states if is_cross_attention else hidden_states if is_cross_attention and past_key_values is not None and is_updated: # reuse k,v, cross_attentions key_states = curr_past_key_values.layers[self.layer_idx].keys value_states = curr_past_key_values.layers[self.layer_idx].values else: key_states = self.k_proj(current_states) value_states = self.v_proj(current_states) key_states = key_states.view(*kv_input_shape).transpose(1, 2) value_states = value_states.view(*kv_input_shape).transpose(1, 2) if past_key_values is not None: # save all key/value_states to cache to be re-used for fast auto-regressive generation cache_position = cache_position if not is_cross_attention else None key_states, value_states = curr_past_key_values.update( key_states, value_states, self.layer_idx, {"cache_position": cache_position} ) # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls if is_cross_attention and isinstance(past_key_values, EncoderDecoderCache): past_key_values.is_updated[self.layer_idx] = True attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.dropout, scaling=self.scaling, output_attentions=output_attentions, **kwargs, ) attn_output = attn_output.reshape(bsz, tgt_len, -1).contiguous() attn_output = self.out_proj(attn_output) return attn_output, attn_weights # Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->M2M100, MBART->M2M100 class M2M100EncoderLayer(GradientCheckpointingLayer): def __init__(self, config: M2M100Config): super().__init__() self.embed_dim = config.d_model self.self_attn = M2M100Attention( embed_dim=self.embed_dim, num_heads=config.encoder_attention_heads, dropout=config.attention_dropout, config=config, ) self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) self.dropout = config.dropout self.activation_fn = ACT2FN[config.activation_function] self.activation_dropout = config.activation_dropout self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim) self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim) self.final_layer_norm = nn.LayerNorm(self.embed_dim) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, output_attentions: bool = False, ) -> torch.Tensor: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. """ residual = hidden_states hidden_states = self.self_attn_layer_norm(hidden_states) hidden_states, attn_weights = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, output_attentions=output_attentions, ) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states residual = hidden_states hidden_states = self.final_layer_norm(hidden_states) hidden_states = self.activation_fn(self.fc1(hidden_states)) hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training) hidden_states = self.fc2(hidden_states) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states if hidden_states.dtype == torch.float16: clamp_value = torch.finfo(hidden_states.dtype).max - 1000 hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) return hidden_states, attn_weights # Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->M2M100, MBART->M2M100 class M2M100DecoderLayer(GradientCheckpointingLayer): def __init__(self, config: M2M100Config, layer_idx: int | None = None): super().__init__() self.embed_dim = config.d_model self.self_attn = M2M100Attention( embed_dim=self.embed_dim, num_heads=config.decoder_attention_heads, dropout=config.attention_dropout, is_decoder=True, is_causal=True, config=config, layer_idx=layer_idx, ) self.dropout = config.dropout self.activation_fn = ACT2FN[config.activation_function] self.activation_dropout = config.activation_dropout self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) self.encoder_attn = M2M100Attention( self.embed_dim, config.decoder_attention_heads, dropout=config.attention_dropout, is_decoder=True, config=config, layer_idx=layer_idx, ) self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim) self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim) self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim) self.final_layer_norm = nn.LayerNorm(self.embed_dim) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor | None = None, encoder_hidden_states: torch.Tensor | None = None, encoder_attention_mask: torch.Tensor | None = None, past_key_values: Cache | None = None, output_attentions: bool | None = False, use_cache: bool | None = True, cache_position: torch.Tensor | None = None, ) -> torch.Tensor: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. encoder_hidden_states (`torch.FloatTensor`): cross attention input to the layer of shape `(batch, seq_len, embed_dim)` encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. past_key_values (`Cache`): cached past key and value projection states output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. It is used to update the cache in the correct position and to infer the complete sequence length. """ residual = hidden_states hidden_states = self.self_attn_layer_norm(hidden_states) # Self Attention hidden_states, self_attn_weights = self.self_attn( hidden_states=hidden_states, past_key_values=past_key_values, attention_mask=attention_mask, output_attentions=output_attentions, cache_position=cache_position, ) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states # Cross-Attention Block cross_attn_weights = None if encoder_hidden_states is not None: residual = hidden_states hidden_states = self.encoder_attn_layer_norm(hidden_states) hidden_states, cross_attn_weights = self.encoder_attn( hidden_states=hidden_states, key_value_states=encoder_hidden_states, attention_mask=encoder_attention_mask, past_key_values=past_key_values, output_attentions=output_attentions, ) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.final_layer_norm(hidden_states) hidden_states = self.activation_fn(self.fc1(hidden_states)) hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training) hidden_states = self.fc2(hidden_states) hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights, cross_attn_weights) return outputs @auto_docstring class M2M100PreTrainedModel(PreTrainedModel): config: M2M100Config base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["M2M100EncoderLayer", "M2M100DecoderLayer"] _supports_flash_attn = True _supports_sdpa = True _supports_flex_attn = True # Doesn't support `compile` (dynamic control flow). Can be fixed but low usage model _can_compile_fullgraph = False def _init_weights(self, module): super()._init_weights(module) if isinstance(module, M2M100SinusoidalPositionalEmbedding): emb_weights = module.get_embedding( module.num_positions + module.offset, module.embedding_dim, module.padding_idx ) init.copy_(module.weights, emb_weights) class M2M100Encoder(M2M100PreTrainedModel): """ Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a [`M2M100EncoderLayer`]. Args: config: M2M100Config embed_tokens (nn.Embedding): output embedding """ def __init__(self, config: M2M100Config): super().__init__(config) self.dropout = config.dropout self.layerdrop = config.encoder_layerdrop embed_dim = config.d_model self.padding_idx = config.pad_token_id self.max_source_positions = config.max_position_embeddings embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0 self.embed_tokens = M2M100ScaledWordEmbedding( config.vocab_size, embed_dim, self.padding_idx, embed_scale=embed_scale ) self.embed_positions = M2M100SinusoidalPositionalEmbedding( config.max_position_embeddings, embed_dim, self.padding_idx, ) self.layers = nn.ModuleList([M2M100EncoderLayer(config) for _ in range(config.encoder_layers)]) self.layer_norm = nn.LayerNorm(config.d_model) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def forward( self, input_ids: torch.Tensor | None = None, attention_mask: torch.Tensor | None = None, inputs_embeds: torch.Tensor | None = None, output_attentions: bool | None = None, output_hidden_states: bool | None = None, return_dict: bool | None = None, **kwargs, ): r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict # retrieve input_ids and inputs_embeds if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") elif input_ids is not None: self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask) input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] else: raise ValueError("You have to specify either input_ids or inputs_embeds") if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) embed_pos = self.embed_positions(input_ids, inputs_embeds) embed_pos = embed_pos.to(inputs_embeds.device) hidden_states = inputs_embeds + embed_pos hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) attention_mask = create_bidirectional_mask( config=self.config, inputs_embeds=inputs_embeds, attention_mask=attention_mask, ) encoder_states = () if output_hidden_states else None all_attentions = () if output_attentions else None synced_gpus = is_deepspeed_zero3_enabled() or is_fsdp_managed_module(self) for idx, encoder_layer in enumerate(self.layers): if output_hidden_states: encoder_states = encoder_states + (hidden_states,) # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description) dropout_probability = torch.rand([]) skip_the_layer = self.training and dropout_probability < self.layerdrop if not skip_the_layer or synced_gpus: # under fsdp or deepspeed zero3 all gpus must run in sync layer_outputs = encoder_layer( hidden_states, attention_mask, output_attentions=output_attentions, ) hidden_states = layer_outputs[0] if skip_the_layer: layer_outputs = (None, None) if output_attentions: all_attentions = all_attentions + (layer_outputs[1],) hidden_states = self.layer_norm(hidden_states) if output_hidden_states: encoder_states = encoder_states + (hidden_states,) if not return_dict: return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions ) class M2M100Decoder(M2M100PreTrainedModel): """ Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`M2M100DecoderLayer`] Args: config: M2M100Config embed_tokens (nn.Embedding): output embedding """ def __init__(self, config: M2M100Config): super().__init__(config) self.dropout = config.dropout self.layerdrop = config.decoder_layerdrop self.padding_idx = config.pad_token_id self.max_target_positions = config.max_position_embeddings embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0 self.embed_tokens = M2M100ScaledWordEmbedding( config.vocab_size, config.d_model, self.padding_idx, embed_scale=embed_scale ) self.embed_positions = M2M100SinusoidalPositionalEmbedding( config.max_position_embeddings, config.d_model, self.padding_idx, ) self.layers = nn.ModuleList([M2M100DecoderLayer(config, layer_idx=i) for i in range(config.decoder_layers)]) self.layer_norm = nn.LayerNorm(config.d_model) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def forward( self, input_ids: torch.Tensor | None = None, attention_mask: torch.Tensor | None = None, encoder_hidden_states: torch.Tensor | None = None, encoder_attention_mask: torch.Tensor | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.Tensor | None = None, use_cache: bool | None = None, output_attentions: bool | None = None, output_hidden_states: bool | None = None, return_dict: bool | None = None, cache_position: torch.Tensor | None = None, **kwargs, ): r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*): Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. It is used to update the cache in the correct position and to infer the complete sequence length. """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict # retrieve input_ids and inputs_embeds if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") elif input_ids is not None: input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] else: raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing`. Setting `use_cache=False`..." ) use_cache = False # initialize `past_key_values` if use_cache and past_key_values is None: past_key_values = EncoderDecoderCache(DynamicCache(config=self.config), DynamicCache(config=self.config)) batch_size, seq_length = inputs_embeds.size()[:-1] past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0 if cache_position is None: cache_position = torch.arange( past_key_values_length, past_key_values_length + seq_length, device=inputs_embeds.device ) if attention_mask is None and not is_torchdynamo_compiling(): # required mask seq length can be calculated via length of past cache mask_seq_length = past_key_values_length + seq_length attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device) self_attn_cache = ( past_key_values.self_attention_cache if isinstance(past_key_values, EncoderDecoderCache) else past_key_values ) attention_mask = create_causal_mask( config=self.config, inputs_embeds=inputs_embeds, attention_mask=attention_mask, cache_position=cache_position, past_key_values=self_attn_cache, ) encoder_attention_mask = create_bidirectional_mask( config=self.config, inputs_embeds=inputs_embeds, attention_mask=encoder_attention_mask, encoder_hidden_states=encoder_hidden_states, ) # embed positions positions = self.embed_positions(input_ids, inputs_embeds, past_key_values_length) positions = positions.to(inputs_embeds.device) hidden_states = inputs_embeds + positions hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None all_cross_attentions = () if output_attentions else None synced_gpus = is_deepspeed_zero3_enabled() or is_fsdp_managed_module(self) for idx, decoder_layer in enumerate(self.layers): if output_hidden_states: all_hidden_states += (hidden_states,) # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description) dropout_probability = torch.rand([]) skip_the_layer = self.training and dropout_probability < self.layerdrop if not skip_the_layer or synced_gpus: # under fsdp or deepspeed zero3 all gpus must run in sync layer_outputs = decoder_layer( hidden_states, attention_mask, encoder_hidden_states, # as a positional argument for gradient checkpointing encoder_attention_mask=encoder_attention_mask, past_key_values=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) hidden_states = layer_outputs[0] if skip_the_layer: continue if output_attentions: all_self_attns += (layer_outputs[1],) all_cross_attentions += (layer_outputs[2],) hidden_states = self.layer_norm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) if not return_dict: return tuple( v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns, all_cross_attentions] if v is not None ) return BaseModelOutputWithPastAndCrossAttentions( last_hidden_state=hidden_states, past_key_values=past_key_values, hidden_states=all_hidden_states, attentions=all_self_attns, cross_attentions=all_cross_attentions, ) @auto_docstring class M2M100Model(M2M100PreTrainedModel): _tied_weights_keys = { "decoder.embed_tokens.weight": "shared.weight", "encoder.embed_tokens.weight": "shared.weight", } def __init__(self, config: M2M100Config): super().__init__(config) padding_idx, vocab_size = config.pad_token_id, config.vocab_size embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0 self.shared = M2M100ScaledWordEmbedding(vocab_size, config.d_model, padding_idx, embed_scale=embed_scale) self.encoder = M2M100Encoder(config) self.decoder = M2M100Decoder(config) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.shared def set_input_embeddings(self, value): self.shared = value self.encoder.embed_tokens = self.shared self.decoder.embed_tokens = self.shared @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, decoder_input_ids: torch.LongTensor | None = None, decoder_attention_mask: torch.LongTensor | None = None, encoder_outputs: tuple[tuple[torch.FloatTensor]] | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, decoder_inputs_embeds: torch.FloatTensor | None = None, use_cache: bool | None = None, output_attentions: bool | None = None, output_hidden_states: bool | None = None, return_dict: bool | None = None, cache_position: torch.Tensor | None = None, **kwargs, ) -> tuple[torch.Tensor] | Seq2SeqModelOutput: r""" decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) M2M100 uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict if encoder_outputs is None: encoder_outputs = self.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True elif return_dict and not isinstance(encoder_outputs, BaseModelOutput): encoder_outputs = BaseModelOutput( last_hidden_state=encoder_outputs[0], hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, ) # decoder outputs consists of (dec_features, past_key_values, dec_hidden, dec_attn) decoder_outputs = self.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, encoder_hidden_states=encoder_outputs[0], encoder_attention_mask=attention_mask, past_key_values=past_key_values, inputs_embeds=decoder_inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, ) if not return_dict: return decoder_outputs + encoder_outputs return Seq2SeqModelOutput( last_hidden_state=decoder_outputs.last_hidden_state, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, encoder_last_hidden_state=encoder_outputs.last_hidden_state, encoder_hidden_states=encoder_outputs.hidden_states, encoder_attentions=encoder_outputs.attentions, ) @auto_docstring( custom_intro=""" The M2M100 Model with a language modeling head. Can be used for summarization. """ ) class M2M100ForConditionalGeneration(M2M100PreTrainedModel, GenerationMixin): base_model_prefix = "model" _tied_weights_keys = {"lm_head.weight": "model.shared.weight"} def __init__(self, config: M2M100Config): super().__init__(config) self.model = M2M100Model(config) self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False) # Initialize weights and apply final processing self.post_init() @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, decoder_input_ids: torch.LongTensor | None = None, decoder_attention_mask: torch.LongTensor | None = None, encoder_outputs: tuple[tuple[torch.FloatTensor]] | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, decoder_inputs_embeds: torch.FloatTensor | None = None, labels: torch.LongTensor | None = None, use_cache: bool | None = None, output_attentions: bool | None = None, output_hidden_states: bool | None = None, return_dict: bool | None = None, cache_position: torch.Tensor | None = None, **kwargs, ) -> tuple[torch.Tensor] | Seq2SeqLMOutput: r""" decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) M2M100 uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. Example Translation: ```python >>> from transformers import AutoTokenizer, M2M100ForConditionalGeneration >>> model = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_418M") >>> tokenizer = AutoTokenizer.from_pretrained("facebook/m2m100_418M") >>> text_to_translate = "Life is like a box of chocolates" >>> model_inputs = tokenizer(text_to_translate, return_tensors="pt") >>> # translate to French >>> gen_tokens = model.generate(**model_inputs, forced_bos_token_id=tokenizer.get_lang_id("fr")) >>> print(tokenizer.batch_decode(gen_tokens, skip_special_tokens=True)) ``` """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict if labels is not None: if decoder_input_ids is None: decoder_input_ids = shift_tokens_right( labels, self.config.pad_token_id, self.config.decoder_start_token_id ) outputs = self.model( input_ids, attention_mask=attention_mask, decoder_input_ids=decoder_input_ids, encoder_outputs=encoder_outputs, decoder_attention_mask=decoder_attention_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, decoder_inputs_embeds=decoder_inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, ) lm_logits = self.lm_head(outputs[0]) masked_lm_loss = None if labels is not None: # move labels to the correct device to enable PP labels = labels.to(lm_logits.device) loss_fct = CrossEntropyLoss() masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1)) if not return_dict: output = (lm_logits,) + outputs[1:] return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output return Seq2SeqLMOutput( loss=masked_lm_loss, logits=lm_logits, past_key_values=outputs.past_key_values, decoder_hidden_states=outputs.decoder_hidden_states, decoder_attentions=outputs.decoder_attentions, cross_attentions=outputs.cross_attentions, encoder_last_hidden_state=outputs.encoder_last_hidden_state, encoder_hidden_states=outputs.encoder_hidden_states, encoder_attentions=outputs.encoder_attentions, ) __all__ = ["M2M100ForConditionalGeneration", "M2M100Model", "M2M100PreTrainedModel"]